#!/usr/bin/env python3
"""Convergence chain: cos(imatrix@N, imatrix@M) for consecutive checkpoints.

If cos(N, 2N) -> 1.000 the extra data changed nothing the quantiser can see,
i.e. the corpus is saturated at N chunks and everything beyond is wasted time.
"""
import sys, numpy as np
from gguf import GGUFReader

def load(path):
    r = GGUFReader(path)
    v = {t.name[:-len(".in_sum2")]: np.array(t.data, dtype=np.float64).reshape(-1)
         for t in r.tensors if t.name.endswith(".in_sum2")}
    n = None
    for f in r.fields.values():
        if f.name == "imatrix.chunk_count":
            n = f.contents()
    return n, v

paths = sys.argv[1:]
loaded = [load(p) + (p.split("/")[-1],) for p in paths]
loaded.sort(key=lambda x: x[0] or 0)

def cos_stats(a, b):
    common = sorted(set(a) & set(b))
    sims = []
    for k in common:
        x, y = a[k], b[k]
        if x.shape != y.shape:
            continue
        nx, ny = np.linalg.norm(x), np.linalg.norm(y)
        if nx and ny:
            sims.append(float(np.dot(x, y) / (nx * ny)))
    s = np.array(sims)
    return s

print("%-12s %-12s %10s %10s %10s %8s" % ("from", "to", "mean cos", "median", "min", "<0.99"))
print("-" * 68)
for i in range(len(loaded) - 1):
    n1, v1, _ = loaded[i]
    n2, v2, _ = loaded[i + 1]
    s = cos_stats(v1, v2)
    print("%-12s %-12s %10.6f %10.6f %10.6f %8d" % (
        "%d ch" % n1, "%d ch" % n2, s.mean(), np.median(s), s.min(), int((s < 0.99).sum())))

print()
print("against the FINAL imatrix (%d chunks):" % loaded[-1][0])
print("%-12s %10s %10s %10s %8s" % ("checkpoint", "mean cos", "median", "min", "<0.99"))
print("-" * 56)
nf, vf, _ = loaded[-1]
for n, v, _ in loaded[:-1]:
    s = cos_stats(v, vf)
    print("%-12s %10.6f %10.6f %10.6f %8d" % (
        "%d ch" % n, s.mean(), np.median(s), s.min(), int((s < 0.99).sum())))